top of page

Interpreting Machine Learning Models to Build Knowledge: A New Approach Using Decision Trees

My Process of Leveraging Large Language Models to Write a Research Highlight

I tested the ability of ChatGPT to generate a research showcase. The result was a combination of three versions provided by ChatGPT. While ChatGPT was proficient at summarizing the information I provided, it struggled with logical and coherent writing in the context of scientific research.

Initially, I provided ChatGPT with the abstract of my paper and requested a research showcase that would be easily understood by the general public. I asked ChatGPT to use examples to highlight the importance of interpreting machine learning and to emphasize that the method presented in this research could be applied to any research that employs machine learning tools, not just precipitation dataset evaluation. However, the first version of the research showcase almost exclusively focused on precipitation and data uncertainty. This could be due to the fact that I had provided my abstract to ChatGPT, which grabbed its attention more than my instructions. Additionally, the only paragraph about machine learning interpretability was illogical, and it aims to highlight the significance of interpreting machine learning but described a scenario involving inaccurate data.

To illustrate the significance of interpreting machine learning, imagine a scenario where a company wants to develop a model to predict customer behavior based on their browsing history. The model could help the company personalize its marketing strategies and increase sales. However, if the model is trained on biased or incomplete data, it could lead to incorrect predictions and ultimately harm the company's bottom line. By interpreting the decision trees used to build the model, the company can identify and correct the biases and limitations in the data, leading to more accurate predictions.

I then asked ChatGPT to revise the article and emphasized that it should focus more on the interpretation of machine learning and the potential knowledge that people could gain about the natural system. I also suggested that ChatGPT provide examples from hydrology or geoscience fields, rather than marketing strategies (as the last version). ChatGPT apologized for its misunderstanding and provided a new version that focused primarily on the interpretability of machine learning, with no mention of precipitation dataset evaluation.

I expressed my gratitude to ChatGPT for its efforts and issued one final order for ChatGPT to generate a new version by combining two of the research showcase versions. I hoped that it would start with the topic of machine learning and then transition to the field of precipitation. I also attached the conclusion of my research for reference.

The lengthy conclusion seemed to have confused ChatGPT, resulting in a version that delved too deeply into the specifics of my research. It included details about the datasets used, the performance of each dataset, and the factors that contributed to data inaccuracy without contextualizing them within the larger picture of the research. Consequently, I had to piece together the storyline I wanted from mainly the first two versions of ChatGPT's output.

The Final Result of this research highlight:

Machine learning (ML) algorithms have become increasingly popular in recent years due to their ability to analyze large datasets and identify complex patterns. Although ML has been widely used for solving complex problems, the black-box nature of these algorithms has limited their interpretability, and understanding how ML tools arrive at their results can be challenging.

In this context, our research paper, "Disentangling error structures of precipitation datasets using decision trees (," presents a novel method for evaluating the error characteristics of precipitation datasets using interpretable machine learning tools. The study uses binary decision trees to identify the factors affecting the error structures of precipitation datasets and provides insights into the underlying causes of these errors.

The study evaluates three different precipitation datasets across the contiguous United States from 2010 to 2019. The results of the study identified the spatiotemporal errors related to three precipitation products and three important factors tied to product errors, including the distance to the coast, soil types, and DEM.

This study shows how machine learning tools can be used to analyze and interpret complex datasets in hydrology. The decision tree approach utilized in this study provides a new perspective on evaluating precipitation products by considering multiple input variables simultaneously. The results of this study can be used to improve the retrieval algorithms and processing methods of precipitation products, making them more useful for scientific and practical purposes.

The approach presented in this research has several potential applications in hydrology and geoscience. For example, it can help researchers better understand the hydrological cycle by identifying the factors that influence precipitation’s spatial and temporal variability. This knowledge can be used to improve water resource management, drought monitoring, and flood forecasting. Moreover, the method presented in this research can also be applied to other research areas that use machine learning tools, such as urban climate, land cover classification, and hydrological modeling. By interpreting the decision trees used in these applications, researchers can gain insights into the underlying mechanisms that drive the observed patterns and better understand the natural system.

In conclusion, our research showcases the importance of understanding how ML tools can be used to build knowledge in fields such as hydrology, climate science, and geoscience. By using decision trees to disentangle error structures in precipitation datasets, we can improve the accuracy of precipitation products and better predict the environmental change on our planet. Further, our approach to interpreting ML results in the context of natural systems has significant implications for a wide range of fields beyond precipitation evaluation. By understanding what ML tools have learned, we can gain new insights into the complex natural systems that govern our planet and improve our ability to predict and mitigate the impacts of environmental change.

47 views0 comments

Recent Posts

See All


bottom of page